A Qualitative Framework for Reasoning Under Uncertainty

代理人必须在一个不确定的世界中推理和行动。表示这种不确定性的标准方法是使用概率。虽然概率有很多优点，但在某些情况下它似乎是不合适或不必要的。这个项目的长期目标是了解在使用更定性的方法时，我们可以在多大程度上保持概率的好处。重点是一种信念的度量，称为似是而非的度量，其中事件的似是而非只是部分有序空间中的一个元素。可信性度量概括了文献中最著名的方法，包括概率、可能性（基于模糊逻辑）和Dempster-Shafer信念函数。正在进行的调查有四条主线：(1)默认推理——使用似是而非的方式赋予默认的语义，例如“鸟通常（或通常）会飞”。(2)信念动力学——使用合理性来捕捉代理的信念，并定义一个适当的条件反射概念来捕捉这些信念如何随时间变化。(3)定性决策理论——利用合理性作为定性决策方法的基础。(4)复杂性问题——调查使用似是而非的定性推理是否比更多的定量推理更容易。这项研究应该有助于为表示不确定性的强大且易于使用的框架提供工具。希望不确定性下定性和定量推理的结合将比单独使用任何一种方法更有效地为智能体提供推理能力。＊＊＊

Agents must reason and act in an uncertain world. The standard approach to representing this uncertainty is to use probability. While probability has many advantages, there are situations where it seems inappropriate or unnecessary. The long-term goal of this project is to understand to what extent we can maintain the benefits of probability while using more qualitative approaches. The focus is on one measure of belief, called a plausibility measure, where the plausibility of an event is just an element in some partially-ordered space. Plausibility measures generalize the bestknown approaches in the literature, including probability, possibility (based on fuzzy logic), and Dempster-Shafer belief functions. Four main lines of investigation are being pursued: (1) Default reasoning---using plausibility to give semantics to defaults such ``birds typically (or normally) fly''. (2) Belief dynamics---using plausibility to capture an agent's beliefs, and defining an appropriate notion of conditioning to capture how these beliefs change over time. (3) Qualitative decision theory---using plausibility as a basis for a qualitative approach to decision making. (4) Complexity issues---investigating whether qualitative reasoning using plausibility is easier than more quantitative reasoning. This research should help provide tools for a powerful, yet easy-touse, framework for representing uncertainty. It is hoped that a combination of qualitative and quantitative reasoning under uncertainty will be more effective in providing reasoning ability to intelligent agents than either approach used alone. ***